RISE SICS North Collaboration, Sweden
ENTER CASE STUDY
Working with the world leading datacentre research institute, RISE SICS North in Luleå, Sweden, Edgetic initially demonstrated in 2018 a potential of up to 50% in efficiency gains in a typical datacentre using its software methods. These gains are a combination of energy savings and higher utilisation.
Edgetic creates state-of-the-art mathematical models of server behaviour in our advanced hardware testing facility, allowing us to accurately predict the power consumption and heat output of a server when workloads are placed on it.
By combining these models with an intelligent scheduling advisor that “plugs in” to Kubernetes – one of the world’s leading software container technologies - and a set of workloads that stressed a server in a series of ways, the Edgetic technology immediately identified the most efficient resource allocation available. This could be tuned to maximise power reduction, increase in server capacity, or a combination of the two.
The Edgetic intelligent hardware models enabled servers in the datacentre to be set for optimum power consumption but without any of the performance reduction normally associated with the setting, greatly improving instructions-per-watt performance and boosting power reduction gains even further.
(a) Overall efficiency improvements through using Edgetic technology at different levels of data centre utilisation
(b) Breakdown of power improvement and performance improvement at different levels of data centre utilisation (in this case on a 12-core CPU)
Second testing phase
The most recent implementation and tests of Edgetic software on large server clusters at RISE during 2018 and early 2019 have already yielded real gains of 20% in energy and power reduction. Further development is underway to increase the potential gains and a commercial offering is scheduled for delivery mid-2019.
Phase 2 used a Kubernetes cluster running on 215 OCP (Open Compute Project) servers at RISE SICS North. Using an improved model creation system based on feedback and analysis from our earlier RISE testing, the OCP servers were mathematically modelled in our advanced hardware testing facility.
Improvements in the Edgetic scheduling advisor enabled smart scheduling of realistic workloads (such as web servers, databases, and batch processing jobs) without prior knowledge of the workload characteristics.
The Edgetic system intelligently loaded servers throughout the data centre, and resulted in a 20% reduction in power reduction across the whole system.
Power consumption over time: Edgetic solution vs the default Kubernetes scheduler.
Comparison of cluster server utilization per job batch, using the default Kubernetes scheduler and the Edgetic scheduler.